The paste that felt harmless
You have a contract to summarise, a spreadsheet of customers to tidy, a doctor's letter you do not understand. The obvious move is to paste the whole thing in and ask for help. It works beautifully, which is exactly why so many people do it before anyone has thought about whether they should.
This lesson is not about fear. It is about drawing one line clearly enough that you stop having to decide in the moment, because deciding in the moment is where mistakes live.
In plain English
- Training data:
- The text a company uses to improve its AI models. On some consumer plans your conversations may be included.
- Personal data:
- Anything that identifies a living person: names, addresses, account numbers, health details, photos.
- Anonymise:
- Replacing identifying details with neutral labels such as Client A or Person B before you share the text.
- Retention:
- How long a company keeps your conversations after you have finished with them.
Where your words actually go
When you type into an AI tool, the text leaves your device and is processed on someone else's computers. Two things follow from that.
First, training. Many free consumer plans may use your conversations to improve the product, and some let you turn that off in settings. Business, team and enterprise plans usually work differently, with customer content excluded from training by default. Notice the word usually. Terms differ between providers and they change, so the only answer that stays true is the one you check for the specific tool and plan you are on.
Second, and less discussed, training is not the only risk. Conversations are stored for some period. Accounts get compromised. Shared logins mean a colleague can read your history. Data can be requested in a legal dispute. So the practical question is not only whether a model might learn from your text, but whether you would be comfortable with that text sitting in a system you do not control.
Whatever this lesson says, your employer's policy wins. If you have one, read it. If you do not have one, assume the strictest reading until somebody tells you otherwise, and ask the question in writing so the answer exists.
The red list
These do not go into a general purpose AI tool, on any plan, without explicit approval from whoever owns the risk.
Passwords, API keys, access tokens and recovery codes. There is no version of this that is fine. If you have already pasted one, rotate it today.
Customer or client records. Names with contact details, order histories, support tickets containing personal information, anything you hold on behalf of someone else.
Contracts and agreements under confidentiality. The signed NDA usually covers exactly this and most people have never reread it.
Medical, financial and legal details. Yours or anyone else's. Health information and bank details deserve more care than almost anything else on this list.
Unreleased company information. Pricing not yet public, acquisition talks, redundancy plans, security incidents, unannounced products.
Anything about a person who has not consented. A colleague's performance issue, a friend's diagnosis, a candidate's application.
Checkpoint
Credentials, customer records, confidential contracts, medical or financial details and unreleased company information stay out of general AI tools.
The middle path: anonymise and keep working
Most people meet a rule like that and quietly ignore it, because the work still has to get done. So here is the version that survives contact with a real job.
The AI almost never needs the identifying details. It needs the structure, the numbers and the problem. Replace the specifics with labels before you paste, keep a small key on your own machine, and put the real names back in the final document yourself.
โ Weak prompt
Prompt
Summarise this contract and tell me what is risky. [pastes the full signed agreement with company names, addresses, signatories and the fee schedule]
Output
This agreement between the two named parties runs for 24 months and contains an automatic renewal clause...
The output is useful and the input was not yours to share. Everything identifying went into a system you do not control, and most confidentiality clauses cover precisely this.
โ Good prompt
Prompt
Summarise this contract and flag the risky clauses. The parties are Supplier A and Client B, the fee is redacted, and I have removed names and addresses. Focus on termination, renewal, liability and anything unusual for a services agreement.
Output
Clause 8 renews automatically unless notice is given 90 days before the end of the term, which is longer than typical. Clause 14 caps liability at one month of fees...
Identical quality of help, none of the exposure. The model never needed to know who the parties were.
I need to get help with a document that contains confidential details.
Do not ask me to paste the document yet.
First, list the categories of information I should replace before pasting anything:
names, company names, addresses, account and reference numbers, dates that could
identify an event, job titles that are unique enough to identify a person.
Then give me a simple labelling scheme (Client A, Person B, Site C) and remind me
to keep the key in a file on my own computer.
Below is an anonymised version of a [contract / complaint / report].
Names are replaced with labels and figures are marked as REDACTED.
Do not ask me for the real names or numbers.
Task: [summarise in plain English / list the risks / draft a reply].
Where a redacted figure matters to your answer, tell me what to check rather than
guessing at a value.
DOCUMENT:
[paste the anonymised text]
Write a short, non-confrontational message to my manager or IT team asking which
AI tools are approved here, whether we are on a business plan that excludes our
data from model training, and what I am allowed to paste. Make it sound like
someone trying to do the right thing, not someone reporting a problem.
Under 120 words.
Two habits worth keeping
Turn off training in the settings of any consumer tool you use for anything remotely work adjacent. It takes a minute and it is one of the few controls you actually hold.
And before you press paste, ask the one question that covers all of this: would I be relaxed if this text appeared in a screenshot shown to my client, my boss and the person it is about? If the answer is no, anonymise it first. If it is still no, do not paste it at all.
If you have already pasted something you should not have, you are in a very large club. Delete the conversation, rotate any credentials involved, and tell whoever needs to know. Lesson five covers why hidden mistakes are the expensive ones.
๐ Quiz
Question 1 of 4Which of these should never go into a general purpose AI tool?